Index

• 04.06
• 11.06
• 18.06
• 25.06

28.05

h_el

1. order

p

04.06

p1

p2

• accuracy 2nd. > 1st. > 3rd
• pers mostly more accurate than pred in terms of MAE & RMSE

p3

p23

11.06

30 bins of equal population

h_ep

p11

• more accurate than bins of equal length
• in some classes accuracy pred > pers

h_dif

println("30 bins of equal population")

30 bins of equal population
1.order
dif: -0.004 ± 0.114
dif_cls: 0.1 ± 4.1
2.order
dif: -0.003 ± 0.11
dif_cls: 0.6 ± 3.7
3.order
dif: -0.008 ± 0.124
dif_cls: 0.6 ± 3.9

• +1 step pred accuracy 2. > 1. > 3. order

maes

• +2, 3, 4 steps pred accuracy < pers

sca

Athens

maes

18.06

• [x] fixed index error for >= +2 steps
• [x] adjusted prediction scheme for >= +2 steps for 1. - 3. order:
• test MAE in
  • [x] Berlin
  • [ ] Athens

for i in 1:length(pred) 
        obs1 = data_test[i] # 9:00
        pred1 = predict_od(obs1, od, T) # pred 9:15
        pred[i] = pred1
        if steps > 1 # +2
            pred2 = predict_od(pred1, od, T) # pred 9:30
            pred[i] = pred2
            if steps > 2 # +3 
                pred3 = predict_od(pred2, od, T) # pred 9:45
                pred[i] = pred3
                if steps > 3 # +4 
                    pred4 = predict_od(pred3, od, T) # pred 10:00
                    pred[i] = pred4

mae_s

mae2_s

mae3

boxes1

boxes2

First thought for hybrid method

1. classify CMF bins from historical data
2. predict using above three ways
3. compute MAE & RMSE for each class
4. choose corresponding one with min. error for adopted prediction

TODO

• optimize algorithm

25.06

• [x] test Athens

l = @layout [a b]

mae1

mae2

boxes2

boxes2